Retell Commercial Review, Remedy Request, and Account Resolution
Aegean Taxi | 2026-03-19
Prepared for internal decision-making and external stakeholder communication. This document is grounded in the actual implementation effort, live tests, middleware work, and observed product behavior.
Core conclusion: the Retell-based architecture did not prove dependable enough to run Aegean Taxi’s no-live-agent booking workflow as the primary operational control layer.
1. Executive summary
Aegean Taxi invested significant time, product effort, and external tooling around a Retell-based voice-booking implementation. The project did not fail because the business problem was unsound; it failed because the live voice orchestration did not become reliable enough for a high-control dispatch workflow that cannot tolerate lost rides or live-agent dependency.
• Multiple layers worked in isolation: middleware endpoints, Onde direct submission, Railway deployment, the WhatsApp/SMS location-capture flow, and the POI / ride-history resolver work all produced usable assets.
• The end-to-end live call flow remained unstable: wrong greeting order, repeated or skipped questions, false unsupported-area behavior, inconsistent function invocation, and mismatch between local tests and real-call behavior.
• Business consequence: the product did not deliver a commercially safe autonomous booking workflow for this use case.
2. Project objective
Requirement
Expectation
Business goal
Run a no-live-agent AI phone operator for taxi bookings, with deterministic handoff to messaging and fixed Onde integration.
Operational constraints
No lost rides, hard control over booking flow, strong location recovery, and commercial usability on Greek islands with difficult place names and accents.
Expected role of Retell
Act as the voice layer that follows the booking workflow, triggers backend functions at the correct moment, and keeps the call on a reliable script.
3. What did work
• Aegean Taxi direct booking path: Prepare → Confirm was eventually validated directly.
• Middleware concept: the separation between voice, middleware, and Onde was correct.
• Location-capture fallback: WhatsApp/SMS/browser location sharing worked and is the strongest fallback discovered in the project.
• Resolver data layer: POI enrichment and 70k-row ride-history work materially improved the foundation for location matching.
4. What failed in practice
• Deterministic live-call control: Retell did not reliably follow the required question order or tool-calling sequence in real calls.
• Prompt/state reliability: multiple rounds of stricter prompts, state-machine shaping, and configuration changes did not produce stable behavior.
• Operational confidence gap: local tests and direct API tests succeeded more often than real calls, creating false confidence.
• High-risk domain mismatch: this project’s hardest requirement is speech-to-location handling for villas, hotels, beach clubs, and tourist pronunciations—the exact place where nondeterministic voice behavior hurts most.
5. Business impact
• Time cost: substantial management, engineering, and testing time was spent trying to make the stack commercially usable.
• Opportunity cost: the same time could have been applied directly to a deterministic in-house workflow built around telephony ingress, middleware, and messaging completion.
• Commercial risk: continuing to push the same architecture would risk further delay and spend without enough evidence that it can meet the no-live-agent requirement.
6. Aegean Taxi’s position
Based on the implementation effort and observed results, Aegean Taxi does not consider the current Retell-based solution fit for the intended purpose in its tested form. The product may be suitable for lighter receptionist or FAQ roles, but it did not prove dependable enough here as the primary controller of a dispatch-critical booking workflow.
7. Requested commercial resolution
No-cost end-to-end delivery by Retell + Full Refund. Aegean Taxi requests a full refund of fees and related costs paid for this failed implementation effort, as well as a consultant-led remediation at no extra cost to Aegean Taxi, with Retell owning the remaining implementation work end to end until a working system is delivered.
8. Acceptance criteria if Retell chooses remediation
• Retell owns the delivery: consultants configure the agent, finish middleware wiring, validate all platform settings, and remain responsible until live calls pass.
• The target flow is simplified: voice layer asks only minimal intake questions, triggers the location-capture link automatically, and hands completion to middleware / WhatsApp.
• No false unsupported-area responses: supported islands and in-island locations must not be rejected incorrectly.
• No live-agent dependency: the system must complete the booking process without human operators for the agreed test matrix.
• Measured live-call pass test: Retell and Aegean Taxi agree a real-call acceptance matrix in advance and sign off only after it passes.
9. Attachments and evidence available on request
• project review memo and implementation notes
• middleware implementation artifacts
• Onde direct integration evidence
• location-capture webapp proof of function
• POI and ride-history resolver assets
• Retell / Claude setup thread and test history
Submitted by Aegean Taxi
For commercial review and immediate account resolution.
